One special challenge in adaptive information filtering is the problem of extremely sparse data. So it is very important to learn optimal threshold while filtering the input textual stream. In this paper, an algorithm is presented for the threshold optimization. The algorithm learns fast by using few positive samples. Moreover, most of the quantities the algorithm requires can be updated incrementally, so its memory and computational power requirements are low. It also has the merits of effective, robust, and practically useful. Fudan University's adaptive text filtering system used this algorithm for the first time and came in third in all runs of TREC10. Its T10U and T10F are 0.215 and 0.414 respectively.